ADK.Tool.ExitLoop (ADK v0.0.1-alpha.1)

Built-in tool that allows an LLM inside a LoopAgent to break out of the loop.

When called, the tool returns {:exit_loop, reason}, which LlmAgent detects and converts to an event with actions.escalate = true. The LoopAgent sees the escalate flag and stops iterating.

This mirrors Python ADK's escalation mechanism where an agent inside a LoopAgent can signal that the task is complete and the loop should exit.

Usage

Add the exit_loop tool to any LlmAgent that runs inside a LoopAgent:

agent = ADK.Agent.LlmAgent.new(
  name: "worker",
  model: "gemini-flash-latest",
  instruction: "Do the task. When finished, call exit_loop.",
  tools: [ADK.Tool.ExitLoop.tool()]
)

loop = ADK.Agent.LoopAgent.new(
  name: "loop",
  sub_agents: [agent],
  max_iterations: 10
)

The LLM will call exit_loop when it considers the task done, producing an escalation event that stops the loop immediately.

Summary

Functions

exit_loop?(arg1)

Returns true if the given value is an exit_loop signal.

reason(arg)

Extract the reason from an exit_loop signal.

tool()

Create the exit_loop tool as a FunctionTool.

Functions

exit_loop?(arg1)

@spec exit_loop?(term()) :: boolean()

Returns true if the given value is an exit_loop signal.

Examples

iex> ADK.Tool.ExitLoop.exit_loop?({:exit_loop, "done"})
true
iex> ADK.Tool.ExitLoop.exit_loop?({:ok, "result"})
false

reason(arg)

@spec reason({:exit_loop, String.t()}) :: String.t()

Extract the reason from an exit_loop signal.

Examples

iex> ADK.Tool.ExitLoop.reason({:exit_loop, "all done"})
"all done"

tool()

@spec tool() :: ADK.Tool.FunctionTool.t()

Create the exit_loop tool as a FunctionTool.

The returned tool, when called by the LLM, returns {:exit_loop, reason}. LlmAgent intercepts this signal and emits an escalation event.

Examples

iex> tool = ADK.Tool.ExitLoop.tool()
iex> tool.name
"exit_loop"